Source-linked AI summary
Accelerated Discovery of 3D Printing Materials Using Data-Driven Multi-Objective Optimization
Timothy Erps, Michael Foshey, Mina Konaković Luković, Wan Shou, Hanns Hagen Goetzke, Herve Dietsch, Klaus Stoll, Bernhard von Vacano, Wojciech Matusik
TL;DR
Random or intuition-led sampling of conflicting material objectives may take months or years to uncover optimal regions. The paper couples multi-objective Bayesian optimization with a semi-automated fabrication workflow and identifies 12 Pareto-optimal formulations after 30 iterations, expanding measured performance-space coverage.
Problem
Random or intuition-led sampling of conflicting material objectives may take months or years to uncover optimal regions of the design space.
Source: Multi-objective optimization with… · p. 5Method
The workflow uses multi-objective Bayesian optimization to propose ratios of six primary inks and couples the search with semi-automated formulation preparation and testing.
Source: Base ingredients and material… · p. 3, Multi-objective optimization with… · p. 5, Mixing · p. 11Results
After 150 samples, 12 formulations formed the Pareto set; performance-space coverage increased 584% for compression modulus and toughness after 30 iterations.
Source: Discovered materials with optimal… · pp. 6–7Takeaways & Limitations
The system can be applied to optimizing materials for other performance objectives and manufacturing processes.
Source: Discussion · p. 10The search was limited to human-chosen inks known to print, and jet-valve results do not directly apply to commercial printing processes.
Source: Discussion · pp. 9–10
Abstract
from arXiv · showhide
Additive manufacturing has become one of the forefront technologies in fabrication, enabling new products impossible to manufacture before. Although many materials exist for additive manufacturing, they typically suffer from performance trade-offs preventing them from replacing traditional manufacturing techniques. Current materials are designed with inefficient human-driven intuition-based methods, leaving them short of optimal solutions. We propose a machine learning approach to accelerate the discovery of additive manufacturing materials with optimal trade-offs in mechanical performance. A multi-objective optimization algorithm automatically guides the experimental design by proposing how to mix primary formulations to create better-performing materials. The algorithm is coupled with a semi-autonomous fabrication platform to significantly reduce the number of performed experiments and overall time to solution. Without any prior knowledge of the primary formulations, the proposed methodology autonomously uncovers twelve optimal composite formulations and enlarges the discovered performance space 288 times after only 30 experimental iterations. This methodology could easily be generalized to other material formulation problems and enable completely automated discovery of a wide variety of material designs.
Results
The workflow combines six compatible primary formulations made from a library of eight ingredients, then prepares and tests composite inks. Samples are printed and post-processed so their mechanical properties can be measured.
Base ingredients and material formulations
- Six primary formulations A–F are built from eight ingredients and selected for printability, viscosity, and distinct mechanical properties and compositions.The ingredient library contains one photoinitiator, three diluents, and four oligomers.Source: Base ingredients and material… · p. 3, Figure 2 · p. 4
- The six primaries are automatically dispensed in chosen ratios and homogenized to prepare composite inks suggested by the optimization algorithm.Dispensing uses a six-channel syringe pump, followed by mixing in a dual asymmetric centrifuge.Source: Base ingredients and material… · p. 4
Jet-valve 3D Printing
- Printed samples are post-processed and flattened to reduce variation before compression testing measures their mechanical performance.The authors note that printing-related features such as layer adhesion can affect mechanical properties.Source: Jet-valve 3D Printing · p. 4
Multi-objective optimization with batch evaluations
A multi-objective Bayesian optimization algorithm searches constrained mixtures of six primary inks for trade-offs among toughness, compression modulus, and maximum compression strength. After 30 iterations, the workflow found 12 Pareto-optimal formulations and expanded the sampled performance space.
Multi-objective optimization with batch evaluations
- Because high-C mixtures were too brittle to flatten for testing, the formulation constraints cap C at 0.5 while requiring all six fractions to sum to 1.The other five fractions range from 0 to 1.Source: Multi-objective optimization with… · p. 5
- The algorithm uses multi-objective Bayesian optimization to search for a large-hypervolume Pareto front across toughness, compression modulus, and maximum strength, evaluating four samples per iteration in parallel.The design variables are the ratios of the six primary formulations.Source: Multi-objective optimization with… · p. 5, Figure 3 · p. 6
Discovered materials with optimal trade-offs
- After 150 samples, 12 formulations formed the Pareto set, raising maximum compression strength 70.8% and toughness 50.8% over the six primaries.The Pareto set includes pure primary F; the reported Pareto points were still increasing the hypervolume when the sample budget ended.Source: Discovered materials with optimal… · p. 6, Figure 4 · p. 7
- The optimization increased the Pareto-front hypervolume by a factor of 1.65, while the full pipeline improved it by a factor of 3.25.The respective hypervolumes for the primaries, initial dataset, and final Pareto front were 2.86e+07, 5.64e+07, and 9.32e+07.Source: Discovered materials with optimal… · p. 6
- After 30 iterations, performance-space coverage grew 250% for compression strength and compression modulus, 399% for compression strength and toughness, and 584% for compression modulus and toughness.Figure 5 compares initial coverage spanned by primary formulations with coverage after 30 iterations.Source: Discovered materials with optimal… · p. 7, Figure 5 · p. 8
Discussion
The discovered Pareto formulations show composition-dependent mechanical trade-offs, while the workflow is designed for automation and adaptation beyond 3D printing. Its search is bounded by human-selected printable base inks and a jet-valve process.
Discussion
- The Pareto set offers application-tailored modulus choices for compliant structures and tunable toughness relevant to aerospace, packaging, and medical applications.The highest compression-modulus point is also a primary formulation and Pareto-optimal composition.Source: Discussion · p. 8, Figure 6 · p. 9
- UDMA contributes strongly to high-modulus materials, while high toughness and compression strength are associated with distinct oligomer and diluent composition ranges.High-toughness formulations include specified ranges of urethane-modified acrylate, aliphatic urethane di-acrylate, UDMA, acrylic acid amide, and acrylic acid ester; the highest-performing compression-strength composites have a separately specified mixture.Source: Discussion · p. 8
- The workflow is automation-ready for mixed-polymer performance optimization, and its base compositions could be changed to adapt it to coatings or molding.The authors state that equipment exists for every pipeline step, from mixing to sample machining, allowing full automation.Source: Discussion · p. 9
- The design space is limited to human-chosen inks known to print, potentially excluding combinations outside the base inks.The authors suggest algorithmic base-composition selection or composing from individual chemicals as alternatives.Source: Discussion · pp. 9–10
- The discovered formulations provide optimal trade-offs in compression modulus, compression strength, and toughness, and the system could extend to other objectives and manufacturing processes.The authors describe the system as a foundation for material engineers and polymer chemists to optimize formulations for varied objectives and applications.Source: Discussion · p. 10
Materials
The study uses commercially sourced ingredients to form six primary inks, then assesses the design space against mixing and jet-valve printability constraints. The fabrication pipeline is semi-automated, with some labor still needed between steps.
Materials
- Modified urethane di-acrylate, aliphatic urethane di-acrylate, polyoxyethyleneamine, photoinitiator, diluents, surfactants, and UDMA are sourced from named suppliers.The passage also identifies aliphatic urethane hexa-acrylate as acquired from Sartometer and acrylic acid amide and acrylic acid ester as supplied by Rahn AG.Source: Materials · p. 10
Making 3D printable primary inks
- Potential formulations must meet mixing and jet-valve viscosity and surface-tension limits to avoid leakage, spreading, poor droplet formation, or clogging.The authors report checking that ink rheology remained stable during mixing across the design space.Source: Making 3D printable primary inks · p. 10
Formulation Inks Compositions
- Six primary inks form the design space, each containing oligomer, reactive diluent, photoinitiator, and surfactant with distinct roles in material properties and printing.The oligomer transfers most physical properties; the diluent lowers viscosity and becomes part of the material, the photoinitiator enables UV polymerization, and surfactant reduces surface tension for droplet formation.Source: Formulation Inks Compositions · p. 10
Handling and Experimental Cost
- Dispensing, mixing, printing, post-processing, and testing can be completed without human input, but transferring materials between pipeline steps requires labor.The pipeline was designed to minimize production and testing cost, with step durations listed in Table S1.Source: Handling and Experimental Cost · pp. 10–11
Dispensing
A syringe-pump system meters the primary formulations into a mixing cup, and a dual-axis centrifuge mixer homogenizes them. The mixing step is optimized for speed.
Dispensing
- A six-axis syringe-pump system simultaneously dispenses primary inks in ratios proportional to their contributions, using steady-state flow to handle differences in viscosity and surface tension.Stepper motors compress syringes routed through a manifold to one mixing cup.Source: Dispensing · p. 11
Mixing
- 64.7 seconds at 3000 rpm mixes each formulation in a dual-axis centrifuge that rapidly circulates material within the vessel.The authors use a SpeedMixer to reduce mixing time during preparation.Source: Mixing · p. 11
3D Printing
The printer uses a piezo-actuated jet-valve to dispense inks with varied fluid properties, followed by UV curing layer by layer. Print-head settings are adjusted per formulation to maintain consistent droplets and printability.
3D Printing
- A 300-micron-nozzle piezo-actuated jet-valve opens to dispense pressurized ink, then closes to cut off flow; dispensed quantity depends on opening time, back pressure, and fluid properties.The spring opens the needle valve when the actuator releases it, and the actuator closes it afterward.Source: 3D Printing · p. 11
- Two jet-valve dispensers deposit materials in parallel, while a 365 nm UV LED cures ink and a Cartesian robot positions the print head and LED.A controller interprets the g-code file that controls robot movement and component actuation.Source: 3D Printing · p. 11
- The printer deposits droplets along a vector path to form each layer, then rasters the UV LED over the layer before repeating at the next height.The process continues until the prescribed 3D object has been dispensed.Source: 3D Printing · pp. 11–12
- Slicing software generates print instructions that are post-processed for UV steps and jet-valve operation; 0.5 mm droplet spacing was chosen to avoid voids.The imported samples were four 8 mm by 8 mm cylinders in a grid array.Source: 3D Printing · p. 12
- Syringe pressure, valve opening percentage, and opening time are manually adjusted per formulation to normalize droplet mass, balancing reliable jetting against splashing.Higher settings increase droplet size, while settings that are too high can cause splashing on the substrate.Source: 3D Printing · p. 12
Post Processing of Printed Samples
The methods standardize printed samples and streamline compression testing, then process stress–strain data into mechanical performance measures.
Post Processing of Printed Samples
- Post-curing at 60°C with 365 nm/405 nm UV exposure completes polymerization, and machining flattens samples to reduce height variation.The samples are baked for one hour in a post-curing oven.Source: Post Processing of Printed Samples · p. 12
Mechanical Performance Testing of Printed Samples
- Compression testing uses 8 mm-diameter, approximately 400 mm³ samples and reduces testing time by avoiding tensile testing’s precision-positioning step.A 3 mm/minute compression rate maintained consistency; samples are measured three times, and their properties are averaged for optimization.Source: Mechanical Performance Testing of… · pp. 12–13
Performance Data Analysis
- Stress–strain analysis trims initial-loading noise and post-failure data, then estimates modulus by filtering sliding-window fits and takes peak stress and strain from the trimmed curve.The automated trimming and modulus calculations were compared with a mechanical engineer’s calculations on initial datasets.Source: Performance Data Analysis · p. 13
Design Space and Initial Samples
The design space is a constrained six-component simplex, initially sampled before a Gaussian-process-based multi-objective algorithm proposes batches of experiments.
Design Space and Initial Samples
- The six formulation ratios lie on a 5-simplex: five range from 0 to 1, one from 0 to 0.5, and all sum to 1.Ratios are limited to two-decimal increments to match dispensing resolution.Source: Design Space and Initial Samples · p. 13
- Twenty-five randomly generated samples uniformly cover the design space as far as possible, alongside five primary formulations for 30 initial samples.The random points are generated from a symmetric Dirichlet distribution.Source: Multi-Objective Optimization Algorithm · p. 13, Design Space and Initial Samples · p. 13
Multi-Objective Optimization Algorithm
- Independent Gaussian processes model each objective, and Thompson sampling balances exploration of unseen regions with exploitation of strong-performing regions.Source: Multi-Objective Optimization Algorithm · p. 13
- The algorithm predicts a Pareto front and set, then selects four samples with the largest expected hypervolume improvement for parallel evaluation in each iteration.Source: Multi-Objective Optimization Algorithm · p. 14
Algorithm Parameters
The implementation specifies Gaussian-process settings, an NSGA-II solver configuration, and a fixed reference point for monitoring hypervolume improvement.
Algorithm Parameters
- Each objective’s Gaussian process uses maximum-a-posteriori hyperparameter estimates and a Matérn 5/2 kernel.The authors cite Bradford et al. for the hyperparameter estimation approach.Source: Algorithm Parameters · p. 14
- NSGA-II uses a population of 100 for 100 generations, with constraints ensuring proposed mutation points satisfy the linear sum-to-one condition.The linear equality constraint is on the six formulation ratios.Source: Algorithm Parameters · p. 14
- A fixed reference point monitors hypervolume improvement and uses the minimum observed value for each objective in the initial dataset.Source: Algorithm Parameters · p. 14
Supplementary Materials
The supplementary materials describe processing-step timing, primary-ink compositions, fabrication equipment, dispensing settings, and optimization metrics.
Supplementary Materials
- Table S1 reports average duration and standard deviation for each sample-fabrication processing step.Source: Table S1 · p. 18
- Table S2 gives the weight-percentage formulations of the six primary inks.Source: Table S2 · p. 18
- Figures S1 and S2 depict the dual asymmetric centrifuge mixer and jet-valve printer used in sample fabrication.Figure S2 identifies parallel dispensers, the UV LED, and the print platform.Source: Figure S1 · p. 19, Figure S2 · p. 19
- Figure S3 compares dispensing droplets across syringe pressure, valve opening pressure, and valve opening time settings.Each setting series varies one parameter while holding the other stated settings constant.Source: Figure S3 · p. 20
- Figure S4 identifies three performance objectives parsed from compression stress–strain data, while Figure S5 illustrates hypervolume over Pareto-optimal points using an origin reference.Source: Figure S4 · p. 20, Figure S5 · p. 21